Chapter 2

CONTROLLABILITY

2.1 Reachable Set and Controllability

Suppose we have a linear system described by the state equation
ẋ = Ax + Bu (2.1)
x(0) = x0
Consider the following problem. For a given vector x in Rn , does there exist a time t, 0 < t < ∞ and
a piecewise continuous control input u, defined on [0, t], such that the solution of (1) x(t) = x? We shall
refer to this problem as the reachability problem.
If for x0 = 0, there exists an input u which solves the reachability problem, we then say that the state
x is reachable from 0 at time t. We refer to such an input u as the control which transfers the state from
the origin to x at time t. Denote the set of states reachable from 0 by R0 . If R0 = Rn , then every state is
reachable from 0 by suitable control.
One motivation for studying reachability is the interception or rendezvous problem. Suppose you know
the trajectory of some vehicle and you want to maneuver your own vehicle to intercept it. If R0 = Rn , the
interception problem is solvable.
To prepare for the discussion later in the chapter, let us give a slightly more abstract but compact
description of the reachability problem. Using the variation of parameters formula, x is reachable from 0
at time t if there exists an input u such that
Z t
x= eA(t−τ ) Bu(τ )dτ (2.2)
0

Fix the time t. Let U be the space of all input signals defined on [0 t]. Define the linear operator L
which maps U to Rn by Z t
Lu = eA(t−τ ) Bu(τ )dτ (2.3)
0
Recall that if A is an m × n matrix (a linear transformation) mapping Rn into Rm , the range of A,
denoted R(A), is the set {y ∈ Rm | y = Ax, for some x ∈ Rn }. Analogously, R(L), the range of L, is
defined to be the set {x ∈ Rn | x = Lu, for some u ∈ U }. Using this formulation, x is reachable from 0 if
and only if x ∈ R(L).
There is nothing particularly special about the initial state x0 = 0. In fact, if R0 = Rn , then every
state x is reachable from any initial state x0 . To see this, write the variation of parameters formula for the
solution of (2.1) in the form Z t
x(t) − eAt x0 = eA(t−τ ) Bu(τ )dτ (2.4)
0

27
28 CHAPTER 2. CONTROLLABILITY

We see that x reachable from x0 is equivalent to x − eAt x0 is reachable from 0.

We now define controllability for the system described by (2.1).
Definition: The system (2.1) is said to be controllable at time t if for every vector x ∈ Rn and every
initial condition x0 , there exists an input u which transfers the state from x0 to x at time t.
In view of the preceding discussion, studying the controllability property is equivalent to studying
whether R0 = Rn .
Example: Consider the following network

i1 i2
R1 R2
+
vi R3
-
+ +
C1 v1 v2 C2
- -

Figure 2.1: An RC bridge circuit

It is straightforward to verify, using Kirchhoff’s current law, that v1 , v2 satisfy the following differential
equations:
v1 − v2 vi − v1
C1 v̇1 = − +
R3 R1
v1 − v2 vi − v2
C2 v̇2 = +
R3 R2
h iT
In state space form, denoting v = v1 v2 , we have
   
−( C11R1 + 1
C1 R3 )
1
C1 R3
1
v̇ =  v +  C1 R1  vi
1
C2 R3 −( C21R2 + 1
C2 R3 )
1
C2 R2

1 1 1 1
Put α1 = C1 R1 , α2 = C2 R2 , β1 = C1 R3 , β2 = C2 R3 . We can write the differential equation in the form
   
−(α1 + β1 ) β1 α
v̇ = v + 1 vi (2.5)
β2 −(α2 + β2 ) α2

Notice that if α1 = α2 , i.e., C1 R1 = C2 R2 , the difference v1 − v2 satisfies

d
(v1 − v2 ) = −(α1 + β1 + β2 )(v1 − v2 )
dt
which is a homogeneous equation with no input. In this case, we cannot manipulate v1 and v2 arbitrarily,
so that the reachable set is not the whole R2 .
To study the reachability problem, we introduce 2 important matrices arising from the system (2.1).
2.1. REACHABLE SET AND CONTROLLABILITY 29

(i) For t > 0, define the controllability Gramian at time t

Z t
T
Wt = eAτ BB T eA τ dτ (2.6)
0
Z t
T
= eA(t−s) BB T eA (t−s) ds (2.7)
0

Finding Wt requires computation of eAt and integration.

Example:
 
−1 1
A =
0 −2
 
s+2 1
−1
0 s+1

s + 1 −1
(sI − A)−1 = =
0 s+2 (s + 1)(s + 2)
" #
1 1  1 1 1 
s+1 (s+1)(s+2) s+1 s+1 − s+2
= 1 = 1
0 s+2
0 s+2

e−t e−t − e−2t

 
At −1 −1
e = L (sI − A) =
0 e−2t
 
0
B =
1
Z t
e−τ e−τ − e−2τ e−τ
   
0 0
Wt = [0 1] dτ
0 0 e−2τ e−τ − e−2τ e−2τ
1
Z t  −τ
e−τ − e−2τ
 
e 0 0
= dτ
0 0 e−2τ e−τ − e−2τ e−2τ
Z t  −2τ
− 2e−3τ + e−4τ e−3τ − e−4τ

e
= dτ
0 e−3τ − e−4τ e−4τ
 1
(1 − e−2t ) − 23 (1 − e−3t ) + 14 (1 − e−4t ) 31 (1 − e−3t ) − 14 (1 − e−4t )

= 2
1 −3t ) − 1 (1 − e−4t ) 1 −4t )
3 (1 − e 4 4 (1 − e

Note that Wt is a symmetric n × n matrix.

(ii) Let CAB denote the n × nm matrix

CAB = [B AB A2 B · · · An−1 B] (2.8)

CAB is called the controllability matrix. For m = 1, CAB is a square matrix. For m > 1, CAB is
a rectangular, wide matrix.

Recall that the rank of a p × q matrix A is the number of independent columns of A, which is the same
as the number of independent rows of A. Note that rank(CAB ) ≤ n. If rank(CAB ) = n, we say CAB has
full rank.
Let us first illustrate the computation of CAB in some examples.
30 CHAPTER 2. CONTROLLABILITY

Example 1: For the circuit described by Figure 2.1, we have

 
α1 −(α21 + α1 β1 ) + α2 β1
CAB =  
2
α2 −(α2 + α2 β2 ) + α1 β2

Note that
det CAB = (α1 α2 + α1 β2 + α2 β1 )(α1 − α2 )
so that CAB is singular (equivalently CAB does not have full rank) if and only if α1 = α2 .
Example 2: Consider the following system of 2 carts connected by a spring
y1 y2
u2
K
u1
M1 M2
0

Figure 2.2: A coupled 2-cart system

Newton’s law gives

M1 ÿ1 = −K(y1 − y2 ) + u1
M2 ÿ2 = −K(y2 − y1 ) + u2
h iT
Let us recast the equations in state space form. Define x = y1 ẏ1 y2 ẏ2 . Then
    
0 1 0 0 x1 0 0
     
K
0   x2   M11
 −K
0 0  u1
   
 M1 M1
ẋ =   +  
 0 0 0 1   x3   0 0  u2
    
    
K −K 1
M2 0 M2 0 x4 0 M2

The controllability matrix for this system is given by

 
1 −K K
 0 0 M1 0 0 0 M12 M1 M2 
 1 −K K


 M 0 0 0 M12 M1 M2 0 0 
CAB =  1


 0 1 K −K
 0 0 M2 0 0 M1 M2 M22


 
1 K −K
0 M2 0 0 M1 M2 M22
0 0

B AB A2 B A3 B

Note that CAB has 4 linearly independent columns so that it is full rank.
Before we discuss the main result of Chapter 2, let us review some geometric ideas from linear algebra.
Suppose V is a subspace of Rn . The orthogonal complement of V, denoted by V ⊥ , is the set of all vectors
w such that wT v = 0 for all v ∈ V.
2.1. REACHABLE SET AND CONTROLLABILITY 31

Example: Suppose V is the x − y plane, a subspace of R3 . Then V ⊥ is the z-axis.

Some important properties of orthogonal complements are:
1. V ⊥ is a subspace.
2. The only vector that lies in both V and V ⊥ is the zero vector.
3. Every vector x ∈ Rn can be uniquely decomposed as x = u + w with u ∈ V and w ∈ V ⊥ . We refer
to the decomposition as Rn is the direct sum of V and V ⊥ , written as Rn = V ⊕ V ⊥ .
4. (V ⊥ )⊥ = V
5. Let V and W be subspaces of Rn . V = W if and only if V ⊥ = W ⊥ , and V ⊂ W if and only if
W ⊥ ⊂ V ⊥.
The following is the main result of this chapter. It shows how CAB and Wt feature in the solution of
the reachability problem.

Theorem 2.1: R(L) = R(Wt ) = R(CAB ).

Proof: The theorem will be proved using the following 2 steps.

(1) Prove R(L) = R(Wt ).
(2) Prove R(CAB ) = R(Wt ).

Step 1: Prove R(L) = R(Wt ).

Using property (5) of orthogonal complements, proving R(L) = R(Wt ) is equivalent to proving
R(L)⊥ = R(Wt )⊥ .
(a) R(L)⊥ ⊂ R(Wt )⊥ : Let v ∈ R(L)⊥ . Then for all u ∈ U,
Z t
v T
eA(t−τ ) Bu(τ )dτ = 0
0

This is true for all u if and only if

v T eAτ B = 0, all τ, 0 ≤ τ ≤ t.
But then Z t
Tτ
v T Wt = v T eAτ BB T eA dτ = 0
0
so that v ∈ R(Wt )⊥ .
(b) R(Wt )⊥ ⊂ R(L)⊥ : Let v ∈ R(Wt )⊥ . Then, in particular,
v T Wt v = 0.
But
Z t
T T Tτ
v Wt v = v eAτ BB T eA dτ v
0
Z t
= k v T eAτ B k2 dτ
0

Thus vT W tv = 0 holds if and only if v T eAτ B = 0, 0 ≤ τ ≤ t. This implies v ∈ R(L)⊥

32 CHAPTER 2. CONTROLLABILITY

This completes Step 1 of the proof.

Note that in the proof, we have also shown that, for t > 0, v T eAτ B = 0, 0 ≤ τ ≤ t if and only if
T
v Wt = 0.
Step 2: Prove R(CAB ) = R(Wt )
First note that we can express eAt in the form

eAt = ϕ0 (t)I + ϕ1 (t)A + ... + ϕn−1 (t)An−1 (2.9)

for certain functions {ϕi (t)}. An outline of the proof of this result is given in the problem sets. Intutively,
we can see that this is a consequence of the Cayley-Hamilton Theorem, which states:
If an n × n matrix A has characteristic polynomial p(s) given by

p(s) = sn + p1 sn−1 + · · · + pn−1 s + pn

then
p(A) = An + p1 An−1 + · · · + pn−1 A + pn I = 0
Cayley-Hamilton Theorem implies Ak , k ≥ n is expressible as a linear combination of Aj , 0 ≤ j ≤ n − 1.
This means that the infinite series that defines eAt should be expressible in the form of (2.9).
We can now carry out the proof of Step 2. We first show

v T CAB = 0

if and only if for t > 0

v T eAτ B = 0 for all τ, 0 ≤ τ ≤ t
Suppose v T CAB = 0. Then

v T eAτ B = v T [ϕ0 (τ )I + ... + ϕn−1 (τ )An−1 ]B

 
ϕ0 (τ )I
 ϕ1 (τ )I 
= v T [B AB ... An−1 B]  ..
 

 . 
ϕn−1 (τ )I
= 0

Conversely, suppose v T eAτ B = 0 for all 0 ≤ τ ≤ t. Setting τ = 0 gives

vT B = 0

For k = 1, . . . , n − 1, take the kth derivative of v T eAτ B with respect to τ and evaluate the result at τ = 0.
This gives successively

v T AB = 0
v T A2 B = 0
..
.
v T An−1 B = 0

so that
v T CAB = 0
2.1. REACHABLE SET AND CONTROLLABILITY 33

Now in the proof of Step 1, we have shown that

v T eAτ B = 0, 0 ≤ τ ≤ t

holds if and only if

v T Wt = 0
holds. This allows us to conclude that
v T CAB = 0
if and only if
v T Wt = 0
Therefore v ∈ R(CAB )⊥ if and only if v ∈ R(Wt )⊥ . Hence R(CAB )⊥ = R(Wt )⊥ , which is equivalent to
R(CAB ) = R(Wt ). This completes the proof of Step 2 and the proof of the Theorem.
Observe that since R(CAB ) is independent ot t, Theorem 2.1 implies that the reachable set and reach-
ability properties are in fact independent of t. We can therefore talk about controllability without refer-
ence to t. Furthermore, note that for an n × n matrix M , R(M ) = Rn if and only if M is nonsingular,
and for an n × nm matrix N , R(N ) = Rn if and only rank(N ) = n. Combining these observations and
the Theorem, we obtain the following
Corollary: The following statements are equivalent:

(a) R0 = Rn

(b) Wt is nonsingular for any t > 0

(c) CAB has rank n (full rank)

Since the linear system (2.1) is controllable if and only if R0 = Rn , the equivalence of statements (a)
and (c) of the Corollary can be restated as the following important theorem.

Theorem 2.2: The linear system (2.1) is controllable if and only if Rank [B AB · · · An−1 B] = n.

Theorem 2.2 allows us to check the controllability property using given data A and B. It is a very
important result.
Based on the Theorem 2.2, we are justified in saying:
The pair (A, B) is controllable if Rank[B AB · · · An−1 B] = n.
Suppose x1 is reachable at time t from x0 . It is not difficult to write down explicitly he control that
achieves the transfer from x0 to x1 at time t. In fact, a control input that achieves the transfer can be
verified to be given by
T
u(τ ) = B T eA (t−τ ) ξ (2.10)
where ξ is the solution of Wt ξ = x1 − eAt x0 . If (A, B) is controllable, Wt−1 exists, and (2.10) can be written
explicitly as
T
u(τ ) = B T eA (t−τ ) Wt−1 (x1 − eAt x0 ) (2.11)
Note also that if (A, B) is controllable, the state transfer can be accomplished for any t > 0, no matter
how small, since Wt is nonsingular for any t > 0. However, if t is small, Wt−1 will be large, and the control
input to achieve the transfer will be large too.
34 CHAPTER 2. CONTROLLABILITY

2.2 Some Properties of Controllability

In this chapter, we discuss 2 properties of controllability: invariance under a change of basis and invariance
under state feedback. We also introduce the controllable canonical form for single input systems, which
will be very useful for the pole assignment problem discussed in the next chapter.

(1). Invariance under change of basis:

Recall that if x is a state vector, so is V −1 x for any nonsingular matrix V . In fact, if we let z = V −1 x,

ż = V −1 x
= V −1 AV z + V −1 Bu

so that with z as the state vector, the system matrices change from (A, B) to (V −1 AV, V −1 B). We refer
to this as a change of basis because if we let the columns of V form a new basis for Rn , z is then the
representation of x in this new basis (See the Appendix to Chapter 2 for more details on change of basis).
We call the transformation V featured in the change of basis a similarity transformation.

Theorem 2.3: (A, B) is controllable if and only if (V −1 AV, V −1 B) is controllable for every nonsingular
V.

Proof:

CV −1 AV,V −1 B = [V −1 B V −1 AV V −1 B · · · ]
= V −1 [B AB · · · ]
= V −1 CAB

Since
Rank(V −1 CAB ) = Rank(CAB )

the result is proved.

(2). Invariance under state feedback:

A control law of the form

u(t) = −Kx(t) + v(t) (2.12)

with v(t) a new input, is referred to as state feedback. The closed-loop system equation is given by

ẋ = (A − BK)x(t) + Bv(t)

It is an important property that controllability is also unaffected by state feedback.

Theorem 2.4: (A, B) is controllable if and only if (A − BK, B) is controllable for all K.
A proof of Theorem 2.4 is discussed in the problem sets. Intuitively, the result seems reasonable since
the state feedback law (2.12) can be reversed by setting v = Kx + u. Therefore, the ability to control the
system using the input v should be the same as that of using input u.
2.3. DECOMPOSITION INTO CONTROLLABLE AND UNCONTROLLABLE PARTS 35

2.3 Decomposition into Controllable and Uncontrollable Parts

If the pair (A, B) is not controllable, there is a particular basis in which the controllable and uncontrollable
parts are displayed transparently. We illustrate the choice of basis and the computation involved using the
circuit example in Section 2.1.

Example: We know from Section 2.1, that the circuit described by Figure 1 is not controllable if R1 C1 =
R2 C2 . Using the notation from Section 2.1, and setting α = R11C1 = R21C2 , we can rewrite (2.5) as
   
−(α + β1 ) β1 α
v̇ = v+ v (2.13)
β2 −(α + β2 ) α i

The controllability matrix is given by

α −α2
 
CAB =
α −α2
T
so that s1 = 1 1 is a basis for R(CAB ). Pick a second vector s2 so that s1 and s2 form a basis for R2 ,

 T
for example, s2 = 0 1 . Set
 
  1 0
V = s1 s2 =
1 1
We then have
   
1 0 −(α + β1 ) β1 1 0
V −1 AV =
−1 1 β2 −(α + β2 ) 1 1
  
1 0 −α β1
=
−1 1 −α −(α + β2 )
 
−α β1
=
0 −(α + β1 + β2 )

and     
−1 1 0 α α
V B= =
−1 1 α 0
Note that the change of basis produces the following system equation

ż = Ãz + B̃u

with
 
−1 Ã11 Ã12
à = V AV = (2.14)
0 Ã22
 
−1 B̃1
B̃ = V B = (2.15)
0

The part corresponding to (Ã11 , B̃1 ), with state component z1 , is controllable, while the part corre-
sponding to the Ã22 , with state component z2 , is uncontrollable, since the z2 equation is decoupled from
z1 , and z2 is unaffected by u. In general, whenever CAB is not full rank, we can find a basis so that in the
new basis, A and B take the form given in (2.14) and (2.15), respectively, with (Ã11 , B̃1 ) controllable. The
procedure is:

1. Find a basis for R(CAB ). Denote the vectors in this basis by {v1 , v2 , · · · , vq }, where q = rank(CAB ).
36 CHAPTER 2. CONTROLLABILITY

2. Select an additional n − q linearly independent vectors vq+1 · · · vn so that {v1 , v2 , · · · .vn } form a basis
for Rn . Define the matrix V by
 
V = v1 v2 · · · vn

3. Compute à = V −1 AV and B̃ = V −1 B. à will take the form (2.14) and B̃ will take the form (2.15).

For more details, the appendix to Chapter 2 contains a formal proof that the above procedure yields Ã
and B̃ as described.

2.4 The PBH Test for Controllability

While we normally check controllability by determining the rank of CAB , there is an alternative useful test
for controllability, referred to as the PBH test. We state the result as a theorem.

Theorem 2.5 (PBH Test):

(A, B) is controllable if and only if Rank[A − λI B] = n for all eigenvalues λ of A.
Proof: The statement is equivalent to:

(A, B) is not controllable if and only if rank[A − λI B] < n for some eigenvalue λ of A.

(i) We first prove rank[A − λI B] < n implies (A, B) is not controllable. Suppose rank[A − λI B] < n
for some eigenvalue λ, possibly complex. Thus there exists a vector x, possibly complex such that

x∗ [A − λI B] = 0

where x∗ is complex conjugate transpose of x. This results in

x∗ A = λx∗

and
x∗ B = 0
Then
x∗ AB = λx∗ B = 0
and
x∗ Ak B = λk x∗ B = 0
Thus x∗ [B AB...An−1 B] = 0 so that [Re x∗ ][B AB...An−1 B] = 0 and (A, B) is not controllable.

(ii) We now show (A, B) not controllable implies rank[A − λI B] < n for some eigenvalue λ of A.
Assume (A, B) is not controllable. By the results of Section 2.3, we can find a nonsingular V so that
(V −1 AV, V −1 B) = (Ã, B̃), where à is of the form (2.14) and B̃ is of the form (2.15), with Ã11 q × q
and Ã22 n − q × n − q. Now note that for any λ an eigenvalue of Ã22 ,

Rank(Ã22 − λI) < n − q

Hence  
Ã11 − λI Ã12 B̃1
[Ã − λI B̃] = Rank <n
0 Ã22 − λI 0
2.4. THE PBH TEST FOR CONTROLLABILITY 37

Next observe that

 
−1 V 0
V [A − λI B] = V −1 [AV − λV B]
0 I
= [V −1 AV − λI V −1 B]
= [Ã − λI B̃]
 
V 0
Since V −1 and are both nonsingular,
0 I

Rank[Ã − λI B̃] = Rank[A − λI B] < n

for λ an eigenvalue of Ã22 . Finally, note that by the structure of Ã, an eigenvalue of Ã22 is an
eigenvalue of Ã. Since a change of basis does not change the eigenvalues, λ is also an eigenvalue of
A. This concludes the proof that (A, B) not controllable implies there exists an eigenvalue of A, λ,
such that rank[A − λI B] < n.

It is useful to note that Rank[A−λI B] = n for all eigenvalues λ of A if and only if Rank[A−λI B] = n
for all complex numbers λ. This is because for λ not an eigenvalue of A, Rank(A − λI) = n. We refer to
the eigenvalue which causes Rank[A − λI B] < n as an uncontrollable eigenvalue. We will see in Chapter
3 that such an eigenvalue is in some sense fixed and not movable.
The PBH test is especially useful for checking when a system is not controllable. We give a couple of
examples to illustrate its use.
Example 1: Consider the following (A, B) pair
   
−3 2 1
A= B=
1 −2 1
This corresponds to the bridge circuit example withe the following parameter values: α1 = α2 = 1, β1 = 2,
and β2 = 1. We know that since α1 = α2 , this system is not controllable. We can also verify that the
controllability matrix is  
1 −1
CAB =
1 −1
which is singular. Let us apply the PBH test. First det(sI − A) = s2 + 5s + 4 so that the eigenvalues are
−4 and −1. For the eignenvalue −1, PBH test gives
 
  −2 2 1
A − (−1)I B =
1 −1 1

which has rank 2. On the other hand, for the eigenvalue −4, PHB test gives
 
  1 2 1
A − (−4)I B =
1 2 1

which has rank 1. By the PBH test, the system is not controllable and −4 is an uncontrollable eigenvalue.
Example 2: Consider the following (A, B) pair
   
0 1 0 0 0
−1 0 1 0 −1
A= 0 0
 B= 
0 1 0
2 0 −2 0 2
38 CHAPTER 2. CONTROLLABILITY

By direct expansion along the first row of (sI − A), we find

 
s −1 0 0
1 s −1 0 
det(sI − A) = det  
0 0 s −1
−2 0 2 s
 
1 −1 0
= s2 (s2 + 2) + det  0 s −1
−2 2 s
= s2 (s2 + 2) + s2 = s4 + 3s2

By inspection, we immediately see that with the eigenvalue 0, Rank[A B] < 4. By the PBH test, this
(A, B) pair is not controllable.

2.5 Controllable Canonical Form for Single-Input Controllable Systems

For single input systems, there is a special form of system matrices for which controllability always holds.
This special form is referred to as the controllable canonical form. Using a lower case b to indicate explicitly
that the input matrix is a column vector for a single input system, the controllable canonical form is given
by
 
0 1 ··· 0 0
 .. ..
.

A= . (2.16)
 

 0 0 ··· 0 1 
−αn −αn−1 · · · −α1

 
0
 0 
..
 
b= (2.17)
 
 . 

 0 
1

It is easy to verify that the controllability matrix for this pair (A, b) always has rank n, regardless of
the values of the coefficients αj . Hence the name controllable canonical form. An A matrix taking the
form given in (2.16) is referred to as a companion form matrix. It is straightforward to show that the
characteristic polynomial of the companion form matrix is given by

det(sI − A) = sn + α1 sn−1 + · · · + αn

The description of the controllable realization in Section 1.7 reflects in effect the properties of the pair
(A, b) given by (2.16) and (2.17).
It will be seen in the next chapter that for applications to pole assignment in single-input systems, the
controllable canonical form is particularly convenient for control design. To prepare for that discussion,
we now show that if (A, b) is controllable, but not in controllable canonical form, we can always find a
similarity transformation V so that (V −1 AV, V −1 b) will be in controllable canonical form.
2.5. CONTROLLABLE CANONICAL FORM FOR SINGLE-INPUT CONTROLLABLE SYSTEMS 39

Consider the matrix

 
1 0 ... 0
 α1 
.
 

α .. 
2
= [An−1 b An−2 b · · · b] 
 
V  ..


 . 
 
 αn−2 0 
αn−1 αn−2 . . . α1 1
 
αn−1 αn−2 · · · α1 1
 αn−2 · · · α1 1 0 
= [b · · · An−1 b]  ...
 

... 1 0 0



 α1 ... 0 ··· 0 


1 0 ··· 0 0
= [v1 · · · vn ]

By controllability, V −1 exists, so that its columns v1 , ..., vn form a basis of Rn .

Note that

v1 = An−1 b + α1 An−2 b + ... + αn−1 b

v2 = An−2 b + α1 An−3 b + . . . + αn−2 b
..
.
vn−1 = Ab + α1 b
vn = b

and that

Av1 = An b + ... + αn−1 Ab + αn b − αn b

= −αn b by the Cayley-Hamilton Theorem
= −αn vn
Av2 = v1 − αn−1 vn
..
.
Avn = vn−1 − α1 vn

Thus the matrix representation of A with respect to the basis {v1 ...vn } looks like
 
0 1 ··· 0 0
 .. ..
.

[A]v =  .
 

 0 0 ··· 0 1 
−αn −αn−1 ··· −α1

Similarly, the vector b looks like  

0
 0 
..
 
 = [b]v
 

 . 
 0 
1
40 CHAPTER 2. CONTROLLABILITY

But [A]v and [b]v are then related to the original matrices through

[A]v = V −1 AV
[b]v = V −1 b

so that they are related by a similarity transformation. Thus the new system z(t) = V −1 x(t) will satisfy
an equation of the form
ż = Ac z + bc u
with (Ac , bc ) in controllable canonical form.
APPENDIX TO CHAPTER 2 41

Appendix to Chapter 2: Change of Basis and System Decomposition

Let us first quickly review the operation of a change of basis in linear algebra and how it affects the
representation of vectors and linear transformations.

Let E = {e1 , . . . , en } and V = {v1 , . . . , vn } be two bases in Rn . Any vector x ∈ Rn can be represented
with respect to either the basis E or the basis V. Thus we can write
n
X n
X
x= ξi ei = ηi vi
i=1 i=1

The coefficient ξi is the ith coordinate of x with respect to the E basis. To emphasize this dependence on
the basis, we write
[x]E = [ξ1 ξ2 . . . ξn ]T
Now let A be a linear transformation mapping Rn to Rn . Its action on on the basis vector ei can be
represented by
Xn
Aei = aki ek
k=1

The coefficients {aij } are then the ijth element of the matrix reprentation of A with respect to the basis
E. We write this as [A]E = {aij }. Similarly, if
n
X
Avi = βki vk
k=1

then [A]V = {βij }.

We now describe the operation of a change of basis. Suppose we wish to change the basis for Rn from
E to V. Define the linear transformation P by
n
X
P ei = vi = αki ek
k=1

so that {αij } is the matrix of P in the basis E. Then

n n n X
n n
" n #
X X X X X
x= ξi ei = ηi P ei = ηi αki ek = αki ηi ek
i=1 i=1 k=1 i=1 k=1 i=1

Hence we have
n
X
ξk = αki ηi
i=1

which can be written as

[x]E = [P ]E [x]V
or
[x]V = [P ]−1
E [x]E
42 CHAPTER 2. CONTROLLABILITY

For linear transformations, we have

Xn n
X n
X n
X
Avi = βki vk = βki P ek = βki αjk ej
k=1 k=1 k=1 j=1

but also
n
X n
X n
X
Avi = AP ei = A αki ek = αki ajk ej
k=1 k=1 j=1
On comparing the two representations for Avi , we get
X n n
X
αjk βki = ajk αki
k=1 k=1
In matrix notation, this corresponds to
[P ]E [A]V = [A]E [P ]E
so that

[A]V = [P ]−1
E [A]E [P ]E

We are now ready to describe the details of system decomposition for an uncontrollable pair (A, B),
described in Section 2.3.
Let Rank[B AB...An−1 B] = q < n, and let {v1 ...vq } be a basis for Range[B AB...An−1 B] = M. We
can pick additional basis vectors {vq+1 , ..., vn } so that {v1 , · · · , vq , vq+1 , · · · , vn } form a basis for Rn . Now
observe that by the Cayley-Hamilton Theorem, AM ⊂ M. Hence for 1 ≤ i ≤ q,
q
X
Avi = αji vj
j=1

for some scalars αji , j = 1, ..., q. With respect to the basis V = {v1 , ..., vn }, A takes the form
 
Ã11 Ã12
[A]V = Ã =
0 Ã22
Similarly the ith column of B,
q
X
bi = βji vj
j=1
Thus, we obtain  
−1 Ã11 Ã12
V AV = Ã =
0 Ã22
and  
−1 B̃1
V B = B̃ =
0
where Ã11 is a q × q matrix and B̃1 is a q × m matrix. We claim that (Ã11 , B̃1 ) is controllable. For,
n−1
 
B̃1 Ã11 B̃1 . . . Ã11 B̃1
Rank[B̃ ÃB̃ ... Ãn−1 B̃] = Rank =q
0 0 ... 0
q+k
But by the Cayley-Hamilton theorem, for k ≥ 0, Ã11 B̃1 is linearly dependent on Ãj11 B̃1 , j = 0, ..., q − 1.
Hence
q−1
Rank[B̃1 Ã11 B̃1 ... Ã11 B̃1 ] = q
so that (Ã11 , B̃1 ) is controllable.